Liquid AI releases LFM3 family of six ~2B-parameter multimodal models built on a shared LFM3-1.7B-Base backbone

Liquid AI released six open models called LFM3, all built from one shared base model (LFM3-1.7B-Base: 24 layers, a mix of convolution, sliding-window and global attention, trained on 22T tokens, 256K context). They cover text (1.7B), vision with video input (VL-2B), audio (Audio-2B), a first omni model (Omni-2B) and two "decision" models (d1-2B, d1-omni-2B). The models run in the Hugging Face transformers library, and the release adds new vision and audio encoders and an audio tokeniser called LiquidCodec. Liquid says the text model beats its earlier LFM2.5-2.6B with 35% fewer parameters, and that the vision model lifts its 62-benchmark average from 57.6 to 71.4.

Developers building for phones and other devices now have a family of small models, one backbone covering text, vision and audio, that they can download. Liquid says the text model fits in about 1 GB at 4-bit and can be fine-tuned on a single GPU, which would lower the cost of building narrow, specialised tools. Teams already evaluating small on-device models have a new candidate to test against their own tasks. The wider field is not changed: this is a new generation from one mid-sized lab, and its practical value depends on how it performs outside Liquid's own tests.

Every benchmark figure comes from Liquid itself and has not been independently checked. The WebArena-Verified jump from 3.02 to 38.42 looks like a hand-picked headline number, and a gain that large against a weak baseline says little about general quality. Phrases like "starting point for specialised, on-device intelligence" are positioning, and the post does not show real-world latency, battery or memory results on actual devices.

1. 01Independent benchmark results (for example community leaderboards or third-party evaluations) for LFM3-1.7B and LFM3-VL-2B, especially whether the WebArena and 62-benchmark gains hold up.
2. 02Real on-device measurements: memory use, speed and quality at 4-bit on phones and laptops, and whether the roughly 1 GB figure holds in practice.
3. 03Adoption signals over the next one to two months: Hugging Face downloads, fine-tuned derivatives, and the licence terms for commercial use.